LLM Manifesto, March 2026 (Subject to Change)
I don't care about AI. I care about outcomes. A process owner's beliefs about agents, autonomy, and where the bottleneck actually moved.
Blog
Thoughts on web development, design, and technology.
SSHler: Managing Multiple Machines Without the Window Chaos
I built a browser-based SSH terminal manager because juggling tmux sessions across four machines via Alt-Tab was making me type commands into the wrong box. Then I started routing LLM agents through it.
proofler: 430 Checks Across Five Libraries, Zero Workarounds
Integration testing five SQLite-backed Python libraries through their public APIs. 16 bugs found and fixed at the source. What the seams between libraries actually look like under chaos.
sqler: The Columnar Baseline
Parts 1–4 measured ORM overhead: sqler vs raw sqlite, both using JSON storage. This post asks a different question. What does the document-oriented architecture itself cost? Equality filter: 11x. Aggregates: 9.5x. JSONL export: 1.0x.
logler: Three Rounds with the Profiler
Hierarchy building from 86 seconds to 332ms (259x). Search from 3.77s to 31ms (121x). Memory from 83 MB to 1 MB (83x). Two measurement bugs caught before they shipped.
qler: The Honest Benchmark
qler vs Celery+Redis — 7 scenarios, 3 rounds, and one embarrassing 12x gap we found by benchmarking ourselves honestly.
sqler: The Final Scorecard
1,725 measurements, 4 scales, 10.5 hours. Bulk insert 0.89x (faster than raw sqlite). FTS ranked 1.00x. Everything else ≤1.15x. One irreducible gap at 1.34x.
sqler: Five Fixes, Five Days
FTS rebuild from 4.65x to 1.03x (benchmark bug). Bulk insert from 1.9x to 0.89x (faster than raw sqlite). FTS ranked from 1.50x to 1.00x (single JOIN). Plus msgspec at 5.1x hydration.
sqler: Where the Overhead Actually Lives
Fair benchmarks pointed at the real bottleneck: Pydantic validation at 1,600ns/row. Bypass it for exports and the 2.8x gap drops to parity. The msgspec question remains open.
sqler: Probably Good Enough Benchmarks
We wrote 22 benchmarks; every one was biased in our favor. An adversarial audit found 18 fairness issues. The rewrite made every number worse but useful.